Flatness Defect Recognition Method of Cold Rolling Strip with a New Stacked Generative Adversarial Network

Abstract

Accurate and rapid recognition of flatness defects is important to produce high‐quality cold rolling strips. First, this study collects the flatness defect images and creates the first image dataset of flatness defect (YSU_CFC_1) according to its characteristics. Then, a flatness defect recognition model of the cold rolling strip with a new stacked generative adversarial network is proposed; it is trained through a recurrent strategy. The generators in this recognition model can generate more real and various fake image samples of flatness defects according to categories. The classifier is used to classify flatness defect images. The recognition results of this model are compared with recognition models established by four convolution neural networks. Results show that the accuracy of this model is the highest, and the accuracy reaches 98.625%, the recognition accuracy of single type defects is ≥95%, and the average recognition time of a single defect image is 24.13 ms, which meets the engineering requirements. Finally, the classification and recognition mechanism of this model is interpretably analyzed, the causes of the recognition error are explored, and the direction of further optimization is provided. The results of this research provide new methods and ideas for the sophisticated detection of flatness defects.